Ambient Noise Probability Density Functions



Characterizing Sources of Noise and Signal in the PDFs

Processing Methods




A new system for analyzing data quality is now available to the seismology community allowing users to evaluate the long-term seismic noise levels for any broadband seismic data channel streaming into the buffer of uniform data (BUD) within the Incorporated Research Institutions for Seismology (IRIS) data management system (DMS). BUD is the IRIS DMS's acronym for the online data cache from which the DMC distributes near-real time miniSEED data holdings prior to formal archiving.

The new noise processing software uses a probability density function (PDF) to display the distribution of seismic power spectral density (PSD) and has been implemented
against the entire continuous data-stream available within the BUD utilizing the QUACK framework. QUACK is the software package at the IRIS DMC, responsible for managing the quality control (QC) of the real-time seismic data flowing into the BUD (see

This noise processing system is unique in that there is no need to screen the data for  earthquakes, system glitches or general data artifacts, as is commonly done in seismic noise analysis. Instead with this new analysis, system transients map into a low-level background probability while ambient noise conditions reveal themselves as high probability occurrances.  In fact, examination of artifacts related to station operation and episodic cultural noise allows us to estimate both the overall station quality and a baseline level of earth noise at each site.

PDF noise plots are useful for characterizing the current and past performance of existing broadband sensors, for detecting operational problems within the recording system, and for evaluating the overall quality of data for a particular station. The advantages of this new approach include:
        1) provides an analytical view representing the true ambient noise levels rather than a simple absolute minimum;
        2) provides an assessment of the overall health of the instrument/station; and,
        3) provides an assessment of the health of recording and telemetry systems.

Below are PDF examples, with some artifacts and signals identified.

HLID - Typical ANSS station ~10km from  Hailey Idaho.
LTX - Lajitas TX. This station was instrumental in the orginal Peterson Low Noise Model, however due to increased cultural noise (0.1-1s, 1-10Hz) the highest probabilty power levels (i.e. mode, black line) are now significantly higher than the LNM. The minimum (red line) will approach the LNM <2% of the time.

Additional features and artifacts observed in the PDFs will be discussed below.  (For a more detailed discussion of the PDFs see McNamara and Buland (2004).)

Back to top

Characterizing Sources of Noise and Signal in the PDFs

Cultural noise

The most common source of seismic noise is from the actions of human beings at or near the surface of the Earth. This is often referred to as “cultural noise” and originates primarily from the coupling of traffic and machinery energy into the earth. Cultural noise propagates mainly as high-frequency surface waves (>1-10Hz,1-0.1s) that attenuate within several kilometers in distance and depth. For this reason cultural noise will generally be significantly reduced in boreholes, deep caves and tunnels. Cultural noise shows very strong diurnal variations and has characteristic frequencies depending on the source of the disturbance.

HLID - automobile traffic along a dirt road only 20 meters from station HLID creates a 20-30dB increase in power at about 0.1 sec period (10Hz). This type of cultural noise is observable in the PDFs as a region of low probability at high frequencies (1-10Hz, 0.1-1s).

ANMO - Borehole station in Albuquerque NM, with very low noise levels. For periods >1s the minimum closely tracks the Peterson LNM.

DWPF - Borehole station at Disney World Florida. Cultural noise levels are high due to human activity and microseism are high due to proximity to the Atlantic ocean.

ISCO - Abandoned mine tunnel station in Idaho Springs CO. Noise levels are low with minimal variations in cultural noise and microseisms.

Our approach differs from many previous noise studies in that we make no attempt to screen the continuous waveforms to eliminate body and surface waves from naturally occurring earthquakes. Earthquake signals are included in our processing because they are generally low probability occurrences even at low power levels (small magnitude events). We are interested in the true noise that a given station will experience, thus we include all signals. For example, including events tells us something about the probability of teleseismic signals being obscured by small local events as well as various noise sources. Large teleseismic earthquakes can produce powers above ambient noise levels across the entire spectrum and are dominated by surface waves >10s, while small events dominate the short period, <1s. Earthquakes are observed in the PDFs as low probability smeared signal at short and long periods.

HLID - Body waves occur as low probabily signal in the 1sec range while surface waves are generally higher power at longer periods.

DWPF - Noise levels are high enough (0.1-1s) that body waves from smaller earthquakes are not readily detectable.

System artifacts

Since we make no attempt to screen waveforms for system transients such as data gaps and sensor glitches, the PDF plots contain numerous system generated artifacts that can be very useful for network quality control purposes. We have attempted to determine the source of several coherent, high power, low probability noise artifacts in the PDF plots. Several artifacts in the PDFs are easily explained and may be useful to the network operator. For example, data-gaps (due to telemetry drop-outs) and automatic mass re-centers (necessitated by “drift” in sensor mass position) are easily identifiable in the PDFs.  Should the probability of mass re-centering and/or telemetry drop-outs drastically increase, a remote network operator could readily diagnose the problem.

HLID  - Automatic mass re-centering and calibration pulses show up as low probability occurances in the PDF.

BINY - Binghamton NY, BHN component. Numerous mass recenters due to unstable sensor pad. We can determine that mass re-centers are automatically issued aproximately 1% of the time.

LTX - Lajitas TX.  Telemetry dropouts are somewhat common (~1-2%).

Back to top

Processing Methods

Power Spectral Density

Employing the algorithm used to develop the USGS Albuquerque Seismological Laboratory (ASL) low noise model (LNM; Peterson, 1993), we compute the power spectral density (PSD) for broadband stations streaming into the IRIS DMC BUD system.

In this algorithm, hour-long, continuous, and over-lapping (50%) time series segments are processed. (N.B. There is no removal of earthquakes, system transients and/or data glitches.) The instrument transfer function is removed from each segment, yielding ground acceleration (for easy comparison to the LNM).  Each hour-long time series is divided into 13 segments, each about 15 minutes long and overlapping by 75%, with each segment processed by:        
        1) removing the mean;
        2) removing the long period trend;
        3) tapering using a 10% sine function; and
        4) transforming using an FFT algorithm (Bendat and Piersol, 1971).

Segments are then averaged to provide a PSD for each one-hour time series segment. 

Probability Density Functions

For each channel, raw frequency distributions are constructed by gathering individual PSDs in the following manner:
       1) binning periods in 1/8 octave intervals; and
       2) binning power in 1 dB intervals.

Each raw frequency distribution bin is then normalized by the total number of PSDs to construct a Probability Density Function (PDF). The probability of occurrence of a given power at a particular period is plotted for direct comparison to the Peterson high and low noise models (HNM, LNM).

We also compute and plot the minimum, mode, and maximum powers for each period bin. A wealth of seismic noise information can be obtained from this statistical view of broadband seismic noise. For a more detailed discussion of the noise processing methods see McNamara and Buland (2004).

Click on the following links for sample PDFs with comments:


or Return to IRIS DMC PDFs

Back to top


Bendat, J.S. and A.G. Piersol (1971). Random data: analysis and measurement procedures. John Wiley and Sons, New York, 407p.

McNamara, D.E. and R.P. Buland, Ambient Noise Levels in the Continental United States, Bull. Seism. Soc. Am., 94, 4, 1517-1527, 2004.

Peterson, J., Observation and modeling of seismic background noise, U.S. Geol. Surv. Tech. Rept., 93-322, 1-95, 1993.

Back to top


The algorithm and initial software were first developed by Dan McNamara and Ray Buland, at the United States Geological Survey (USGS) as a part of the data and network quality control (QC) system for the Advanced National Seismic System (ANSS).  Further development, supported by IRIS through funds it receives from the National Science Foundation (NSF) allowed for Richard Boaz to develop the system for implementation against the BUD dataset. Additonal installation and implementation support was provided by Bruce Weertman and Tim Ahern at the IRIS DMC.

Webpage authors: D. McNamara (USGS ANSS) and R. Boaz (

Back to top